How Criteo optimized and sped up its TensorFlow models by 10x and served them under 5 ms:
When you access a web page, bidders such as Criteo must determine in a few dozens of milliseconds if they want to purchase the advertising space on the page. At that moment, a real-time auction takes place, and once you remove all the communication exchange delays, it leaves a handful of millisecond...
Gespeichert in:
| Beteiligte Personen: | , |
|---|---|
| Körperschaft: | |
| Format: | Elektronisch Video |
| Sprache: | Englisch |
| Veröffentlicht: |
[Erscheinungsort nicht ermittelbar]
O'Reilly Media, Inc.
2020
|
| Ausgabe: | 1st edition. |
| Schlagwörter: | |
| Links: | https://learning.oreilly.com/library/view/-/0636920372547/?ar |
| Zusammenfassung: | When you access a web page, bidders such as Criteo must determine in a few dozens of milliseconds if they want to purchase the advertising space on the page. At that moment, a real-time auction takes place, and once you remove all the communication exchange delays, it leaves a handful of milliseconds to compute exactly how much to bid. In the past year, Criteo has put a large amount of effort into reshaping its in-house machine learning stack responsible for making such predictions-in particular, opening it to new technologies such as TensorFlow. Unfortunately, even for simple logistic regression models and small neural networks, Criteo's initial TensorFlow implementations saw inference time increase by 100, going from 300 microseconds to 30 milliseconds. Nicolas Kowalski and Axel Antoniotti outline how Criteo approached this issue, discussing how Criteo profiled its model to understand its bottleneck; why commonly shared solutions such as optimizing TensorFlow build for the target hardware, freezing and cleaning up the model, and using accelerated linear algebra (XLA) ended up being lackluster; and how Criteo rewrote is models from scratch, reimplementing cross-features and hashing functions using low-level TF operations in order to factorize as much as possible all TensorFlow nodes in its model. Prerequisite knowledge A basic understanding of how TensorFlow and TensorFlow Serving work Experience optimizing TensorFlow models for serving (useful but not required) What you'll learn Understand how to optimize a TensorFlow model before serving it online Discover how to profile a TensorFlow model with a complex preprocessing architecture Learn how and when to replace feature columns with custom cross-features and hashing functions to factorize and drastically reduce the number of nodes in the model This session is from the 2019 O'Reilly TensorFlow World Conference in Santa Clara, CA. |
| Beschreibung: | Online resource; Title from title screen (viewed February 28, 2020) |
| Umfang: | 1 Online-Ressource (1 video file, approximately 38 min.) |
| Format: | Mode of access: World Wide Web. |
Internformat
MARC
| LEADER | 00000cgm a22000002c 4500 | ||
|---|---|---|---|
| 001 | ZDB-30-ORH-050573659 | ||
| 003 | DE-627-1 | ||
| 005 | 20240228121008.0 | ||
| 006 | m o | | | ||
| 007 | cr uuu---uuuuu | ||
| 008 | 200324s2020 xx ||| |o o ||eng c | ||
| 035 | |a (DE-627-1)050573659 | ||
| 035 | |a (DE-599)KEP050573659 | ||
| 035 | |a (ORHE)0636920372547 | ||
| 035 | |a (DE-627-1)050573659 | ||
| 040 | |a DE-627 |b ger |c DE-627 |e rda | ||
| 041 | |a eng | ||
| 082 | 0 | |a E VIDEO | |
| 100 | 1 | |a Kowalski, Nicolas |e VerfasserIn |4 aut | |
| 245 | 1 | 0 | |a How Criteo optimized and sped up its TensorFlow models by 10x and served them under 5 ms |c Kowalski, Nicolas |
| 250 | |a 1st edition. | ||
| 264 | 1 | |a [Erscheinungsort nicht ermittelbar] |b O'Reilly Media, Inc. |c 2020 | |
| 264 | 2 | |a Boston, MA |b Safari. | |
| 300 | |a 1 Online-Ressource (1 video file, approximately 38 min.) | ||
| 336 | |a zweidimensionales bewegtes Bild |b tdi |2 rdacontent | ||
| 337 | |a Computermedien |b c |2 rdamedia | ||
| 338 | |a Online-Ressource |b cr |2 rdacarrier | ||
| 500 | |a Online resource; Title from title screen (viewed February 28, 2020) | ||
| 520 | |a When you access a web page, bidders such as Criteo must determine in a few dozens of milliseconds if they want to purchase the advertising space on the page. At that moment, a real-time auction takes place, and once you remove all the communication exchange delays, it leaves a handful of milliseconds to compute exactly how much to bid. In the past year, Criteo has put a large amount of effort into reshaping its in-house machine learning stack responsible for making such predictions-in particular, opening it to new technologies such as TensorFlow. Unfortunately, even for simple logistic regression models and small neural networks, Criteo's initial TensorFlow implementations saw inference time increase by 100, going from 300 microseconds to 30 milliseconds. Nicolas Kowalski and Axel Antoniotti outline how Criteo approached this issue, discussing how Criteo profiled its model to understand its bottleneck; why commonly shared solutions such as optimizing TensorFlow build for the target hardware, freezing and cleaning up the model, and using accelerated linear algebra (XLA) ended up being lackluster; and how Criteo rewrote is models from scratch, reimplementing cross-features and hashing functions using low-level TF operations in order to factorize as much as possible all TensorFlow nodes in its model. Prerequisite knowledge A basic understanding of how TensorFlow and TensorFlow Serving work Experience optimizing TensorFlow models for serving (useful but not required) What you'll learn Understand how to optimize a TensorFlow model before serving it online Discover how to profile a TensorFlow model with a complex preprocessing architecture Learn how and when to replace feature columns with custom cross-features and hashing functions to factorize and drastically reduce the number of nodes in the model This session is from the 2019 O'Reilly TensorFlow World Conference in Santa Clara, CA. | ||
| 538 | |a Mode of access: World Wide Web. | ||
| 650 | 0 | |a Machine learning | |
| 650 | 0 | |a Artificial intelligence | |
| 650 | 4 | |a Apprentissage automatique | |
| 650 | 4 | |a Intelligence artificielle | |
| 650 | 4 | |a artificial intelligence | |
| 650 | 4 | |a Electronic videos | |
| 700 | 1 | |a Antoniotti, Axel |e VerfasserIn |4 aut | |
| 710 | 2 | |a Safari, an O'Reilly Media Company. |e MitwirkendeR |4 ctb | |
| 966 | 4 | 0 | |l DE-91 |p ZDB-30-ORH |q TUM_PDA_ORH |u https://learning.oreilly.com/library/view/-/0636920372547/?ar |m X:ORHE |x Aggregator |z lizenzpflichtig |3 Volltext |
| 912 | |a ZDB-30-ORH | ||
| 935 | |c vide | ||
| 951 | |a BO | ||
| 912 | |a ZDB-30-ORH | ||
| 049 | |a DE-91 | ||
Datensatz im Suchindex
| DE-BY-TUM_katkey | ZDB-30-ORH-050573659 |
|---|---|
| _version_ | 1831287161118261249 |
| adam_text | |
| any_adam_object | |
| author | Kowalski, Nicolas Antoniotti, Axel |
| author_corporate | Safari, an O'Reilly Media Company |
| author_corporate_role | ctb |
| author_facet | Kowalski, Nicolas Antoniotti, Axel Safari, an O'Reilly Media Company |
| author_role | aut aut |
| author_sort | Kowalski, Nicolas |
| author_variant | n k nk a a aa |
| building | Verbundindex |
| bvnumber | localTUM |
| collection | ZDB-30-ORH |
| ctrlnum | (DE-627-1)050573659 (DE-599)KEP050573659 (ORHE)0636920372547 |
| dewey-raw | E VIDEO |
| dewey-search | E VIDEO |
| edition | 1st edition. |
| format | Electronic Video |
| fullrecord | <?xml version="1.0" encoding="UTF-8"?><collection xmlns="http://www.loc.gov/MARC21/slim"><record><leader>03586cgm a22004692c 4500</leader><controlfield tag="001">ZDB-30-ORH-050573659</controlfield><controlfield tag="003">DE-627-1</controlfield><controlfield tag="005">20240228121008.0</controlfield><controlfield tag="006">m o | | </controlfield><controlfield tag="007">cr uuu---uuuuu</controlfield><controlfield tag="008">200324s2020 xx ||| |o o ||eng c</controlfield><datafield tag="035" ind1=" " ind2=" "><subfield code="a">(DE-627-1)050573659</subfield></datafield><datafield tag="035" ind1=" " ind2=" "><subfield code="a">(DE-599)KEP050573659</subfield></datafield><datafield tag="035" ind1=" " ind2=" "><subfield code="a">(ORHE)0636920372547</subfield></datafield><datafield tag="035" ind1=" " ind2=" "><subfield code="a">(DE-627-1)050573659</subfield></datafield><datafield tag="040" ind1=" " ind2=" "><subfield code="a">DE-627</subfield><subfield code="b">ger</subfield><subfield code="c">DE-627</subfield><subfield code="e">rda</subfield></datafield><datafield tag="041" ind1=" " ind2=" "><subfield code="a">eng</subfield></datafield><datafield tag="082" ind1="0" ind2=" "><subfield code="a">E VIDEO</subfield></datafield><datafield tag="100" ind1="1" ind2=" "><subfield code="a">Kowalski, Nicolas</subfield><subfield code="e">VerfasserIn</subfield><subfield code="4">aut</subfield></datafield><datafield tag="245" ind1="1" ind2="0"><subfield code="a">How Criteo optimized and sped up its TensorFlow models by 10x and served them under 5 ms</subfield><subfield code="c">Kowalski, Nicolas</subfield></datafield><datafield tag="250" ind1=" " ind2=" "><subfield code="a">1st edition.</subfield></datafield><datafield tag="264" ind1=" " ind2="1"><subfield code="a">[Erscheinungsort nicht ermittelbar]</subfield><subfield code="b">O'Reilly Media, Inc.</subfield><subfield code="c">2020</subfield></datafield><datafield tag="264" ind1=" " ind2="2"><subfield code="a">Boston, MA</subfield><subfield code="b">Safari.</subfield></datafield><datafield tag="300" ind1=" " ind2=" "><subfield code="a">1 Online-Ressource (1 video file, approximately 38 min.)</subfield></datafield><datafield tag="336" ind1=" " ind2=" "><subfield code="a">zweidimensionales bewegtes Bild</subfield><subfield code="b">tdi</subfield><subfield code="2">rdacontent</subfield></datafield><datafield tag="337" ind1=" " ind2=" "><subfield code="a">Computermedien</subfield><subfield code="b">c</subfield><subfield code="2">rdamedia</subfield></datafield><datafield tag="338" ind1=" " ind2=" "><subfield code="a">Online-Ressource</subfield><subfield code="b">cr</subfield><subfield code="2">rdacarrier</subfield></datafield><datafield tag="500" ind1=" " ind2=" "><subfield code="a">Online resource; Title from title screen (viewed February 28, 2020)</subfield></datafield><datafield tag="520" ind1=" " ind2=" "><subfield code="a">When you access a web page, bidders such as Criteo must determine in a few dozens of milliseconds if they want to purchase the advertising space on the page. At that moment, a real-time auction takes place, and once you remove all the communication exchange delays, it leaves a handful of milliseconds to compute exactly how much to bid. In the past year, Criteo has put a large amount of effort into reshaping its in-house machine learning stack responsible for making such predictions-in particular, opening it to new technologies such as TensorFlow. Unfortunately, even for simple logistic regression models and small neural networks, Criteo's initial TensorFlow implementations saw inference time increase by 100, going from 300 microseconds to 30 milliseconds. Nicolas Kowalski and Axel Antoniotti outline how Criteo approached this issue, discussing how Criteo profiled its model to understand its bottleneck; why commonly shared solutions such as optimizing TensorFlow build for the target hardware, freezing and cleaning up the model, and using accelerated linear algebra (XLA) ended up being lackluster; and how Criteo rewrote is models from scratch, reimplementing cross-features and hashing functions using low-level TF operations in order to factorize as much as possible all TensorFlow nodes in its model. Prerequisite knowledge A basic understanding of how TensorFlow and TensorFlow Serving work Experience optimizing TensorFlow models for serving (useful but not required) What you'll learn Understand how to optimize a TensorFlow model before serving it online Discover how to profile a TensorFlow model with a complex preprocessing architecture Learn how and when to replace feature columns with custom cross-features and hashing functions to factorize and drastically reduce the number of nodes in the model This session is from the 2019 O'Reilly TensorFlow World Conference in Santa Clara, CA.</subfield></datafield><datafield tag="538" ind1=" " ind2=" "><subfield code="a">Mode of access: World Wide Web.</subfield></datafield><datafield tag="650" ind1=" " ind2="0"><subfield code="a">Machine learning</subfield></datafield><datafield tag="650" ind1=" " ind2="0"><subfield code="a">Artificial intelligence</subfield></datafield><datafield tag="650" ind1=" " ind2="4"><subfield code="a">Apprentissage automatique</subfield></datafield><datafield tag="650" ind1=" " ind2="4"><subfield code="a">Intelligence artificielle</subfield></datafield><datafield tag="650" ind1=" " ind2="4"><subfield code="a">artificial intelligence</subfield></datafield><datafield tag="650" ind1=" " ind2="4"><subfield code="a">Electronic videos</subfield></datafield><datafield tag="700" ind1="1" ind2=" "><subfield code="a">Antoniotti, Axel</subfield><subfield code="e">VerfasserIn</subfield><subfield code="4">aut</subfield></datafield><datafield tag="710" ind1="2" ind2=" "><subfield code="a">Safari, an O'Reilly Media Company.</subfield><subfield code="e">MitwirkendeR</subfield><subfield code="4">ctb</subfield></datafield><datafield tag="966" ind1="4" ind2="0"><subfield code="l">DE-91</subfield><subfield code="p">ZDB-30-ORH</subfield><subfield code="q">TUM_PDA_ORH</subfield><subfield code="u">https://learning.oreilly.com/library/view/-/0636920372547/?ar</subfield><subfield code="m">X:ORHE</subfield><subfield code="x">Aggregator</subfield><subfield code="z">lizenzpflichtig</subfield><subfield code="3">Volltext</subfield></datafield><datafield tag="912" ind1=" " ind2=" "><subfield code="a">ZDB-30-ORH</subfield></datafield><datafield tag="935" ind1=" " ind2=" "><subfield code="c">vide</subfield></datafield><datafield tag="951" ind1=" " ind2=" "><subfield code="a">BO</subfield></datafield><datafield tag="912" ind1=" " ind2=" "><subfield code="a">ZDB-30-ORH</subfield></datafield><datafield tag="049" ind1=" " ind2=" "><subfield code="a">DE-91</subfield></datafield></record></collection> |
| id | ZDB-30-ORH-050573659 |
| illustrated | Not Illustrated |
| indexdate | 2025-05-05T13:25:31Z |
| institution | BVB |
| language | English |
| open_access_boolean | |
| owner | DE-91 DE-BY-TUM |
| owner_facet | DE-91 DE-BY-TUM |
| physical | 1 Online-Ressource (1 video file, approximately 38 min.) |
| psigel | ZDB-30-ORH TUM_PDA_ORH ZDB-30-ORH |
| publishDate | 2020 |
| publishDateSearch | 2020 |
| publishDateSort | 2020 |
| publisher | O'Reilly Media, Inc. |
| record_format | marc |
| spelling | Kowalski, Nicolas VerfasserIn aut How Criteo optimized and sped up its TensorFlow models by 10x and served them under 5 ms Kowalski, Nicolas 1st edition. [Erscheinungsort nicht ermittelbar] O'Reilly Media, Inc. 2020 Boston, MA Safari. 1 Online-Ressource (1 video file, approximately 38 min.) zweidimensionales bewegtes Bild tdi rdacontent Computermedien c rdamedia Online-Ressource cr rdacarrier Online resource; Title from title screen (viewed February 28, 2020) When you access a web page, bidders such as Criteo must determine in a few dozens of milliseconds if they want to purchase the advertising space on the page. At that moment, a real-time auction takes place, and once you remove all the communication exchange delays, it leaves a handful of milliseconds to compute exactly how much to bid. In the past year, Criteo has put a large amount of effort into reshaping its in-house machine learning stack responsible for making such predictions-in particular, opening it to new technologies such as TensorFlow. Unfortunately, even for simple logistic regression models and small neural networks, Criteo's initial TensorFlow implementations saw inference time increase by 100, going from 300 microseconds to 30 milliseconds. Nicolas Kowalski and Axel Antoniotti outline how Criteo approached this issue, discussing how Criteo profiled its model to understand its bottleneck; why commonly shared solutions such as optimizing TensorFlow build for the target hardware, freezing and cleaning up the model, and using accelerated linear algebra (XLA) ended up being lackluster; and how Criteo rewrote is models from scratch, reimplementing cross-features and hashing functions using low-level TF operations in order to factorize as much as possible all TensorFlow nodes in its model. Prerequisite knowledge A basic understanding of how TensorFlow and TensorFlow Serving work Experience optimizing TensorFlow models for serving (useful but not required) What you'll learn Understand how to optimize a TensorFlow model before serving it online Discover how to profile a TensorFlow model with a complex preprocessing architecture Learn how and when to replace feature columns with custom cross-features and hashing functions to factorize and drastically reduce the number of nodes in the model This session is from the 2019 O'Reilly TensorFlow World Conference in Santa Clara, CA. Mode of access: World Wide Web. Machine learning Artificial intelligence Apprentissage automatique Intelligence artificielle artificial intelligence Electronic videos Antoniotti, Axel VerfasserIn aut Safari, an O'Reilly Media Company. MitwirkendeR ctb |
| spellingShingle | Kowalski, Nicolas Antoniotti, Axel How Criteo optimized and sped up its TensorFlow models by 10x and served them under 5 ms Machine learning Artificial intelligence Apprentissage automatique Intelligence artificielle artificial intelligence Electronic videos |
| title | How Criteo optimized and sped up its TensorFlow models by 10x and served them under 5 ms |
| title_auth | How Criteo optimized and sped up its TensorFlow models by 10x and served them under 5 ms |
| title_exact_search | How Criteo optimized and sped up its TensorFlow models by 10x and served them under 5 ms |
| title_full | How Criteo optimized and sped up its TensorFlow models by 10x and served them under 5 ms Kowalski, Nicolas |
| title_fullStr | How Criteo optimized and sped up its TensorFlow models by 10x and served them under 5 ms Kowalski, Nicolas |
| title_full_unstemmed | How Criteo optimized and sped up its TensorFlow models by 10x and served them under 5 ms Kowalski, Nicolas |
| title_short | How Criteo optimized and sped up its TensorFlow models by 10x and served them under 5 ms |
| title_sort | how criteo optimized and sped up its tensorflow models by 10x and served them under 5 ms |
| topic | Machine learning Artificial intelligence Apprentissage automatique Intelligence artificielle artificial intelligence Electronic videos |
| topic_facet | Machine learning Artificial intelligence Apprentissage automatique Intelligence artificielle artificial intelligence Electronic videos |
| work_keys_str_mv | AT kowalskinicolas howcriteooptimizedandspedupitstensorflowmodelsby10xandservedthemunder5ms AT antoniottiaxel howcriteooptimizedandspedupitstensorflowmodelsby10xandservedthemunder5ms AT safarianoreillymediacompany howcriteooptimizedandspedupitstensorflowmodelsby10xandservedthemunder5ms |